This is just a heads-up for everyone using noflushd, still the only reliable solution to powering down rotational drives. After moving from 5.15 to 5.18 I noticed that dirty writeback cache for any *unstopped* devices (e.g. system root or other SSDs) would accumulate without being written periodically; after some time I found that this was due to noflushd running. It would still shut down the rotational drive(s) it was supposed to suspend, but due to how it works (slightly fugly) and recent changes in the kernel, writeback would no longer happen when noflushd was active. Manually performing sync or stopping noflushd would bring things back to normal. After some debugging I think I have found the reason (kernel >5.15 "cleanups"/changes no longer performing fsync/syncfs on an fd of the raw block device), but need to confirm this further. This might actually be an unintentional kernel bug, though I don't think fsync on a raw bdev was ever _really_ supposed to work. Usually one could just move to using hdparm/sdparm suspend, but - as I just found out - the firmware in my hard drive is batshit buggy and only accepts -B timer values from 1-11; 12+ no longer works and never suspends the drive. Unfortunately noflushd is abandonded on SourceForge, and only available in CVS. I might try to rescue it to Github (help welcome, I have forgotten all about CVS) but in the meantime this is just a warning, since overly large amounts of accumulated dirty bytes can result in unexpectedly large data loss when a power outage or system crash happens. Reproducible: Always Steps to Reproduce: 1. start noflushd on kernel >5.15 2. dirty writeback data is no longer written back to *unstopped* devices
According to Christoph Hellwig fsync on a block device should still work, yet whatever is going on in noflushd is now working differently than before. Potential replacements that are also less invasive/fragile as they do not interfere with the kernel writeback mechanism: - https://sourceforge.net/projects/hd-idle/ Already in portage, but unmaintained. Lots of copies on GH. - https://github.com/adelolmo/hd-idle An up-to-date & maintained rewrite in Go - https://github.com/Piezoid/rust-idle An up-to-date & maintained rewrite in Rust Hope this helps.
Given that there are several alternatives and one in the tree (hd-idle worked well) can we give last rites to noflushd?
The bug has been referenced in the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=e23d31bc532d7f2b15ef09b6fdc321cc1619f9f3 commit e23d31bc532d7f2b15ef09b6fdc321cc1619f9f3 Author: Michał Górny <mgorny@gentoo.org> AuthorDate: 2023-12-31 10:58:23 +0000 Commit: Michał Górny <mgorny@gentoo.org> CommitDate: 2023-12-31 10:59:50 +0000 package.mask: Last rite sys-block/noflushd Bug: https://bugs.gentoo.org/849077 Signed-off-by: Michał Górny <mgorny@gentoo.org> profiles/package.mask | 6 ++++++ 1 file changed, 6 insertions(+)
The bug has been closed via the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=3adc0ce822a378279633e7d7127e0e83ad2c5812 commit 3adc0ce822a378279633e7d7127e0e83ad2c5812 Author: Michał Górny <mgorny@gentoo.org> AuthorDate: 2024-02-02 16:41:15 +0000 Commit: Michał Górny <mgorny@gentoo.org> CommitDate: 2024-02-02 16:41:15 +0000 sys-block/noflushd: Remove last-rited pkg Closes: https://bugs.gentoo.org/849077 Signed-off-by: Michał Górny <mgorny@gentoo.org> profiles/package.mask | 6 ------ sys-block/noflushd/Manifest | 1 - sys-block/noflushd/files/noflushd.confd | 10 ---------- sys-block/noflushd/files/noflushd.rc6 | 30 ------------------------------ sys-block/noflushd/metadata.xml | 11 ----------- sys-block/noflushd/noflushd-2.8-r1.ebuild | 30 ------------------------------ 6 files changed, 88 deletions(-)
